cfcpys and cfcpyd have hardware bugs which mean they truncate denormalized values to zero and convert minus zero to plus zero. A 64-bit shift with a shift count of 0 copies them bitwise. Martin Guy , December 2008 --- gcc-4.3.2/gcc/config/arm/cirrus.md.old 2008-11-21 19:06:37.000000000 +0000 +++ gcc-4.3.2/gcc/config/arm/cirrus.md 2008-12-03 12:13:08.000000000 +0000 @@ -485,7 +485,7 @@ && (GET_CODE (operands[0]) != MEM || register_operand (operands[1], SFmode))" "@ - cfcpys%?\\t%V0, %V1 + cfsh64%?\\t%Z0, %Z1, #0\\t%@ float cfldrs%?\\t%V0, %1 cfmvsr%?\\t%V0, %1 cfmvrs%?\\t%0, %V1 @@ -514,7 +514,7 @@ case 1: return \"stm%?ia\\t%m0, %M1\\t%@ double\"; case 2: return \"#\"; case 3: case 4: return output_move_double (operands); - case 5: return \"cfcpyd%?\\t%V0, %V1\"; + case 5: return \"cfsh64%?\\t%Z0, %Z1, #0\\t%@ double\"; case 6: return \"cfldrd%?\\t%V0, %1\"; case 7: return \"cfmvdlr\\t%V0, %Q1\;cfmvdhr%?\\t%V0, %R1\"; case 8: return \"cfmvrdl%?\\t%Q0, %V1\;cfmvrdh%?\\t%R0, %V1\";